Intended use & limitations

What VetDiff is for

VetDiff is clinical decision support for licensed veterinarians and for veterinary students working under supervision. It helps you build and check a differential diagnosis list, and turns dictated or typed visit notes into a structured SOAP note.

It does not diagnose, prescribe, or replace a physical examination, diagnostic testing, a specialist's opinion, or your clinical judgment. Every diagnosis and treatment decision remains the responsibility of the veterinarian caring for the patient.

It is not designed for emergency triage. In an emergency, treat the patient first; don't wait on a differential.

How a differential is produced

When you submit a case, VetDiff sends the signalment, history, and lab results you entered (including any photos or PDFs) to a large language model — currently Anthropic's Claude — with instructions to return a ranked differential list, the reasoning for each item, suggested next diagnostic steps, and a reference basis.

Two additional sources are included when available: this patient's prior visits at your practice, and up to five similar cases from VetDiff's shared pool of de-identified cases whose diagnoses other veterinarians have confirmed. The shared pool is new and still small, so most differentials today rely mainly on the model's general veterinary knowledge — learned from the veterinary literature and other material it was trained on.

After the differential appears, VetDiff runs a live search restricted to veterinary sources — PubMed and PubMed Central, veterinary journals, the Merck Veterinary Manual, and consensus-guideline bodies such as AAHA, WSAVA, and IRIS — and lists supporting references for the leading differentials. Every reference is checked against the actual search results, so each one links to a real published page; whether a source fully supports a specific point is still for you to judge, so read it before relying on it. Searches use clinical terms only (signalment and condition names), never names or other identifying details.

The separate "reasoning basis" is the model's own summary of the clinical principles behind its reasoning. Any authors or sources it names come from its training and are not checked — rely on the supporting-literature list for verified sources.

Probability tiers (High / Moderate / Low) are qualitative judgments by the model, not calibrated statistical probabilities.

What has — and hasn't — been validated

VetDiff has not yet undergone formal, prospective, or independent validation of its diagnostic accuracy. We will not describe it as validated until it has been.

Our first measure will come from real use: for every case where a veterinarian records the confirmed diagnosis, we track whether that diagnosis appeared in VetDiff's list, and at what rank (top 1, top 3, top 5). We'll publish those figures here once enough confirmed cases exist to be meaningful, along with their limits — confirmations are self-reported and may not represent every kind of case.

We welcome independent evaluation by faculty and clinicians, and can provide evaluation access and sample cases on request.

Known limitations

• The correct diagnosis can be missing from the list, or ranked lower than it should be — especially for rare conditions or unusual presentations.
• The model can state incorrect facts with confidence, including drug doses and reference ranges. Verify doses and treatment details against a current formulary.
• Output quality depends on input quality. Missing negative findings, typos, or an incomplete history all reduce accuracy.
• Values read from photos or PDFs of lab reports can occasionally be misread. Confirm key values against the original report.
• Performance is strongest for dogs and cats in general and emergency practice. Exotic, zoo, large-animal, and specialty cases are less reliable.
• Dictation can mis-transcribe medical terms, names of drugs, and numbers, and a formatted note can occasionally omit or misfile a detail. Review every note before saving it to the medical record.
• VetDiff can't examine the patient. It only knows what you tell it.

Using it safely

• Build your own differential first, then compare it with VetDiff's — use disagreements as prompts to think, not answers.
• Treat every item as a hypothesis to confirm or rule out with examination and testing.
• Don't enter client-identifying information (owner names, addresses, phone numbers). Case text is automatically de-identified before it can enter the shared pool, but the less identifying detail entered, the better.
• Students: follow your school's and teaching hospital's policies, and your supervising clinicians' decisions.

Reporting a problem

If VetDiff produces something wrong, unsafe, or misleading, please tell us: email support@vetdiff.com or use the suggestion box in the app, and include the case date and what went wrong. Every report is reviewed by the veterinarian who built VetDiff, and we aim to respond within two business days.

Confirmed problems lead to changes in how VetDiff prompts or checks the model, and material changes to how VetDiff works are noted on this page with the date.

Diagnoses confirmed from student accounts are kept out of the shared pool, so trainee outcomes never shape differentials for practicing veterinarians.